Predicting flight arrival times with deep learning: A strategy for minimizing potential conflicts in gate assignment

Abstract

Air transportation is frequently disrupted by factors such as weather and air traffic control, making it difficult for flights to strictly adhere to schedules, leading to frequent early arrivals or delays. These disruptions pose challenges to airport operations management, particularly in gate assignments, where potential conflicts and adjustments are often required. Unlike traditional methods that focus on enhancing robustness to reduce conflicts, this study adopts a Predict-then-Optimize (PO) framework, using predicted flight arrival times for gate assignments to avoid the need for robustness-related objectives. In the prediction phase, a CNN-LSTM-Attention deep learning model is developed to predict flight arrival times based on the historical data of a single airport, enhancing data availability and model practicality. In the optimization phase, a bi-objective gate assignment model is constructed, using predicted arrival times instead of scheduled times as input. An epsilon-constrained branch-and-price algorithm is developed to obtain non-dominated Pareto optimal solutions. Analysis using actual operational data from Beijing Capital International Airport shows that the prediction model achieves an accuracy of 93.27% for early arrivals and 83.6% for on-time flights. The epsilon-constrained branch-and-price algorithm outperforms heuristic algorithms in both the quantity and quality of Pareto solutions. Notably, the gate assignment strategy based on predicted arrival times significantly reduces potential conflicts, with a maximum reduction of 25.33% compared to the schedule-based strategy. This study demonstrates that the proposed gate assignment method, based on flight arrival time prediction, effectively mitigates the impact of arrival time uncertainty on gate assignments, providing a new approach to reducing potential conflicts without relying on robustness.